SNR-Adaptive Unified Diffusion for Multi-Task Medical Image Segmentation

The paper introduces UniT-Diff, a single-parameter diffusion segmentation framework that jointly handles semi-supervised learning, unsupervised domain adaptation, and domain generalisation for cardiac images. It addresses conflicting label semantics with an 11-channel task-specific output space, adjusts task conditioning according to the diffusion timestep’s log signal-to-noise ratio, and removes task conditioning for domain-generalisation inputs. The authors report gains over independently trained task-specific baselines on all three benchmarks: +0.87 percentage points on LA, +1.77 on MMWHS, and +0.88 on MNMS.